There is a technique I now call the reference roster, and it fixes the most annoying thing about AI image editing: the face that changes slightly every time you ask for a new background, or the product label that quietly rewrites itself by the third edit. Google's Nano Banana 2.1, released on 6 October 2026, is built to hold subjects steady across rounds of editing. It only does that well if you tell it clearly which image is which.
This guide covers what changed in 2.1, why drift happens, a copy-paste roster prompt for campaign images, a one-change-per-turn editing method, and the limits you should know before you put a 2.1 image in front of a client.
What is Nano Banana 2.1?
Nano Banana 2.1 is Google's everyday image generation and conversational editing model, released on 6 October 2026 as the successor to Nano Banana 2. It accepts up to 14 reference images per request, holds up to 4 characters and 10 objects consistent, and is designed to keep faces and products recognisable across several rounds of edits.
Google's model page lists the key updates: better visual quality at 1K, 2K and 4K, sharper text and infographic layouts, fixed tiling artifacts on very wide or tall images, grounding with Google Web and Image Search, and three thinking levels (minimal, medium and high, with medium as the default).
You can use it without writing code. According to Fello AI's launch coverage, Google rolled it out the same day to the Gemini app, AI Mode in Google Search (tap the banana icon under the search box), Google AI Studio, Flow and Stitch. Developers call it in the Gemini API as gemini-nano-banana-2.1.
On Arena's crowd-voted leaderboard on launch day it ranked fifth for text-to-image and sixth for image editing. That is the highest of any Google image model, but still behind OpenAI's GPT Image 2.5 and GPT Image 2.
Why do faces and products drift between AI image edits?
Drift happens because each edit is a fresh generation, not a pixel patch. The model re-draws the whole image from your words plus whatever images it can see. If the prompt re-describes the subject, references are unlabelled, or the chat has run for many turns, the model blends your instructions with its own guesses and small details shift.
In practice I see three causes again and again:
--- Unlabelled references. You upload a product shot and a model photo, then say "put her holding the bottle". The model has to guess which image is "her" and which is "the bottle", and sometimes it borrows colours or features across them.
--- Over-describing the subject. Writing "a woman with shoulder-length black hair and a round face" alongside a reference photo creates two sources of truth. The text usually wins, and the face drifts towards your description.
--- Long edit chains. Every turn adds a little noise. By the eighth "now make it warmer" request, the label font or a jawline has moved, even though no single step looked wrong.
Google says 2.1's main improvement over Nano Banana 2 is "multi-turn character consistency". The reference caps stay the same as before. What changes is how reliably the model keeps a subject recognisable across rounds, which means your setup matters more than ever.
How do you build a reference roster in Nano Banana 2.1?
A reference roster is a short labelled list at the top of your prompt that gives every uploaded image a number and a single job: subject, product, style or setting. You then refer to images only by those labels and describe the new scene, not the subject. This removes guesswork about which image controls what.
Keep the roster within Google's stated limits: up to 4 characters and up to 10 objects, 14 images in total. In my experience, fewer is better. Two people and two products give cleaner results than pushing the cap.
Choose clean references. A front-facing face with even lighting, or a product shot where the label, logo and shape are fully visible, gives the model the most to anchor on. Cropped, blurry or heavily filtered images give it the least.
Try this prompt (attach 3 images in this order):
REFERENCE ROSTER
Image 1 = SUBJECT A. The person. Keep face, hairstyle, skin texture and body proportions identical.
Image 2 = PRODUCT. The bottle. Keep shape, cap colour, label text and logo position identical. Do not redraw or translate the label.
Image 3 = STYLE ONLY. Use its colour palette and soft afternoon lighting. Do not copy any objects or people from it.
SCENE
SUBJECT A sits at a cafe table by a window in Hong Kong, holding PRODUCT at chest height with the label facing the camera. Relaxed smile, looking slightly off camera.
OUTPUT
4:5 portrait, 2K. Natural photo style, shallow depth of field. No text overlays.
If any roster item cannot be kept identical, tell me which one before generating.
Notice what the prompt does not do: it never describes the person's face or the bottle's design. The roster says "keep identical", and the scene section talks only about pose, place, light and framing.
For a campaign, keep the roster block word-for-word the same and change only the SCENE section for each image. That one habit does more for consistency across a set of five or ten images than any adjective you could add.
How do you edit one part of an image without changing the rest?
Use semantic masking: name the single element to change, describe the replacement, and explicitly tell the model to leave everything else alone. Make one change per turn, and when you need several, restart from the best version with your roster attached rather than stacking edits in a long chat.
Google's image generation guide calls this approach semantic masking, and Google's launch post said 2.1 makes "notable leaps" in mask-based editing. You describe the area in words instead of painting a selection.
Try this prompt (attach the image you want to edit):
Using the provided image, change only the background wall from white tiles to a dark green painted wall with a framed print. Keep the person, the product, the table, the lighting direction and the camera angle exactly the same. Do not change the label on the bottle.
Three rules keep edits clean:
--- One change per turn. "Change the wall and make her jacket red and add a plant" invites the model to re-draw everything. Split it into three turns.
--- Name what must not change. List the two or three things you care most about, especially text and faces.
--- Reset after about four edits. Download the best version, open a new chat, attach it with your original roster images, and continue from there. This clears the accumulated drift.
If you work in Google AI Studio or the API, you can also set the thinking level. Medium is the default; high helps with crowded scenes with many constraints, while minimal is cheaper and fine for simple background swaps. In the Gemini app and AI Mode you do not choose this yourself.
Where does Nano Banana 2.1 still fall short?
Nano Banana 2.1 is better at holding subjects steady, but it still misspells small text, struggles with tiny faces in crowd shots, cannot use web photos of real people for grounding, and stamps every image with an invisible SynthID watermark. Treat it as a fast first-draft tool and check every output before publishing.
--- Text still needs proofreading. Google improved text rendering, but its own Nano Banana 2 documentation warns about spelling and fine details. Check every label, price and Chinese character by eye.
--- Small faces are weak. Consistency is strongest when the subject fills a good part of the frame. A group shot with six small faces is where likenesses slip first.
--- No real people from Image Search. According to Google's documentation, as reported by Fello AI, grounding with Google Image Search cannot use web photos of real people. Use your own licensed photos as references.
--- Watermarked by default. Every image carries Google's invisible SynthID mark. That is useful for transparency, but tell clients up front if their brand guidelines have rules about AI-generated imagery.
--- No 512px size in the API. 2.1 starts at 1K. If you previously generated cheap 512px drafts with Nano Banana 2, which Google shuts down in the API on 29 October 2026, plan for that change.
--- Not the top-ranked model. On Arena's launch-day boards, GPT Image 2.5 and GPT Image 2 still rank higher. For a single hero image, test both before you commit.
Try it now: a 20-minute consistency test
Run one subject and one product through three scenes using the same roster block, then compare the three images side by side. If the face and the label survive all three, your roster works and you can scale it to a full campaign. If not, simplify the references before blaming the model.
--- Minutes 0 to 5: pick one clean portrait you have rights to and one clear product photo. Open the Gemini app or AI Mode and attach both.
--- Minutes 5 to 10: paste the roster prompt above, adjusted to your images. Generate scene one.
--- Minutes 10 to 15: in a new chat, attach the same two images, keep the roster identical, change only the SCENE section to an office, then to an outdoor street.
--- Minutes 15 to 20: put the three images side by side. Check the eyes, hairline, label text and logo position. Write down which item drifted.
Whatever drifted tells you what to fix: a better reference photo, a stricter "keep identical" line, or fewer items on the roster. The skill here is not clever adjectives. It is giving the model one clear job per image and refusing to describe what it can already see. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team. Model details reflect Google's Nano Banana 2.1 model page (updated 6 October 2026) and Fello AI's launch report of 7 October 2026; availability, limits and rankings may change. Related reading: our Midjourney V8.2 Edit Model guide and output contracts for consistent AI results.
Test how sharp your AI skills really are
Reference rosters and one-change edits are the kind of habits that separate casual AI users from power users. Take UD's free AI IQ Test to see where you stand, and when you want these techniques built into a repeatable team workflow, we'll walk you through every step, from tool setup to workflow design and deployment.
